Frank Van Damme a écrit :
> On Monday 22 April 2002 10:36 am, you wrote:
> > You need a cluster here.
> > An extremely powerfull server that can server easily all your clients will
> > always be a single point of failure. Immagine 500 people calling in the
> > same hour your IT stuff because the server went out.
>
> Maybe a decent server with redundant disks, power supplies and even cpu's
> will me more reliable then 6 crap PC boxes.
>
> > You need a cluster.
> > Depending on your budget you could use two IBM S390/Zseries (500000$+ per
> > standard server) or a cluster of independent servers -
> > http://www.idg.net/go.cgi?id=671252 -.
>
> Mainframes aren't made for this. They are extremely bad as calculators and
> for graphical stuff. They are good for io-intensive applications (like
> webserving).
> Didn't know it. Thank You.
> > Go for the architecture you have experience with. I mean even if alphas are
> > powerfull, use only x86 if you are used to use them.
>
> C'mon. If you are serious about a solution for a big environment at least
> take the time to get to know the stuff you will need. btw I have a small old
> sparc here, makes no difference. (I am talking wbout sparcs not alphas,
> because I never same close to an alpha.) The difference with small and big
> sparcs is that there is more stuff is a big one. I went shopping at sun (no
> panic I don't know your credit card number ;) ) and came out at 57500 USD,
> this means: 115 usd clients not included. We're talking about a 4-way machine
> (750 MHz sparc cpu's with 8 MB cache), 8 gigs of ram and 6 36-gig fast scsi
> disks here.
>
>http://store.sun.com/webconfig/BuildConfig.jhtml;$sessionid$32JBHKI5R00FDAMTA1ESRUT5AAAACJ1K#Summary
>
A sun entreprise 450 costs 100 engineer's salaries and an ibm h70 costs
125 here.
I don't know how much is 57500USD (conversion has no meaning when life
costs are very different).
Ther is also a unique sun representative and I heared that a lot of
people (like universities) are
changing from sun because they are not satisfied with
support/maintenance/acquisition prices.
So a cluster could be really less expensive than a big sun.
> You could use (this is how I would proceed):
> > 1 or 2 p133 as your dhcp server (BBC uses 1 dhcp server per 33000 users -
> > http://www.linuxplanet.com/linuxplanet/reports/1176/2/ - and has one ready
> > in case of failure of the first one). The rescue dhcp server could also be
> > the nfs server. On nfs server, where all /home stuff would reside, with
> > RAID5 scsi disks and a raid controller with a lot of ram cache, one
> > dedicated nic per application server 100Mbps (or 1000Mbps if possible but
> > not necessary) and a twisted cat5 connection between the nfs server and
> > each application one with no hub or swich in the middle. People seem to
>
> That makes sense!
>
> > have happily used 200 clients on a single server so I would recommend for
> > your 500 clients using 6 2way x86 application servers (ltsp servers) with
> > 2~4 Go ram per server (we use nt4 servers with 2xPIII500 Mhz cpus, 1Gig
>
> 12-24 GB ram may be what NT needs, but for linux this seems a 'bit' much to
> me allready.
>
> > RAM and RAID5 fast SCSI2 disks per server per 20 clients, citrix inc. says
> > we could server 40 clinets with the same machine). These should have fast
> > scsi swap disks. When run as mosix cluster nodes, the application servers
> > could also load all the / stuff from the nfs server, easing the
> > maintenance.
> > Configure your dhcp server so clients will be divided in 5 or 6 segments
> > with each segment starting X sessions from a different application server.
> > A lot of things allows you to do load balancing (use dns instead of ips:
> > the fastest available app server will respond first; xdmcp chooser: people
> > will see which server have the less users connected and use it; I think
>
> How does that work?
You could set "option x-display-manager appserv.yourcompany.com;" in
your dhcpd.conf instead of
indicating the ip address of the server. This will make the clients do a
dns lookup, wich in turn could
return several ips or a different ip for each request.
The terminal client will then use that ip to connect to the server. If
it is possible to set a timeout
for the client waiting the server to respond, this could make the client
query the next server or you
could more simply put all the ips of your 6 nodes in a circular order
for each group/subnet in the
dhcpd.conf. So if a server is overloaded and doesn't respond, the client
will query the next one.
I am not an expert of the xdmcp nor x protocols so please correct me if
I am saying "divagations - from
fench".
>
>
> > round-robin dns will respond whith a different ip each time it is requested
> > to do a dns lookup and could be used for load balancing) but remember that
> > you are running a cluster and thus all the load will be shared by the
> > different nodes at execution time (the cluster will self manage its load).
> > As people seemed to be distracted and slowed down by the KDE2 interface, I
> > switched to ICEwm (a lot faster and does not make any difference as people
> > use the window manager to load apps with mouse clicks and not to work with
> > it).
>
> The problem with kde's interface is it's responsiveness on slow computers.
> Did you monitor cpu usage? I think the X server still gets away with most of
> the cpu time. Moreover this should have been improved greatly with kde 3.0.
I have an heterogenous set of computers starting from compaq 486/22 with
4/8 megs.
I also don't allow write access on the nfs server, so it was not
reasonable to continue using KDE and I
hided from login konsole.
Fathi Ben Nasr.
>
>
> Frank
> --
> homepage: www.student.kuleuven.ac.be/~m9917684
> jabber (=IM): [EMAIL PROTECTED]
> No part of this copyright message may be reproduced, read or seen,
> dead or alive or by any means, including but not limited to telepathy
> without the benevolence of the author.
_____________________________________________________________________
Ltsp-discuss mailing list. To un-subscribe, or change prefs, goto:
https://lists.sourceforge.net/lists/listinfo/ltsp-discuss
For additional LTSP help, try #ltsp channel on irc.openprojects.net
